Statistical Consulting at Draper Laboratory

Statistical Consulting at Draper Laboratory A Project Report Submitted to the Faculty of the WORCESTER POLYTECHNIC INSTITUTE in partial fulfillment of the requirements for the Degree of Master of Science in Applied Statistics Submitted By: Noelle M. Richard Professor Joseph D. Petruccelli, Advisor August,

Abstract This Master s capstone was conducted in conjunction with Draper Laboratory, a non-profit research and development organization in Cambridge, Massachusetts. During a three month period, the author worked for the Microfabrication Department, assisting with projects related to statistics and quality control. The author gained real-world experience in data collection and analysis, and learned a new statistical software. Statistical methods covered in this report include regression analysis, control charts and capability, Gage R & R studies, and basic exploratory data analysis.

Table of Contents Abstract... Figures... Tables... Introduction... 6 Chapter : Etch Process... Introduction... Goals... Data/Variables... 8 Statistical Analysis... 8 Comments/Recommendations... 6 Chapter : Denton... Introduction... Goals... Statistical Analysis... Comments... 8 Chapter 3: COIST Wafers... 33 Part... 33 Introduction... 33 Goals... 33 Data/Variables... 33 Statistical Analysis... 3 Conclusions... 3 Part... 39 Introduction... 39 Goals... 39 Statistical Analysis... 39 Chapter : Gage R & R Studies... Part : Review of Past Studies... Introduction... Past Procedure... Goals... 3 Statistical Analysis... Comments/Suggestions... 6 Part : New Studies... 6 Goals... 6 Tencor... 6 Tencor P 6... Filmetrics F... Summary... 3 Chapter : Minitab PowerPoint Presentations... Introduction... Data/Variables... Presentations... Conclusions/Critique... Concluding Remarks... 8 Acknowledgements... 8 References... 9

Figures Figure : Etch Process... Figure : Distribution of Cavity Diameter... 9 Figure 3: Normal Quantile Plot of Cavity Diameter... 9 Figure : General Linear Model : Cavity Diameter vs. all other variables... Figure : Residual Plots from General Linear Model... Figure 6: Stepwise Regression with both Forward and Backward Elimination... 3 Figure : Residual Plots from General Linear Model... 3 Figure 8: General Linear Model : Cavity Diameter vs. Horizontal Undercut, Mask Diameter, Etch Time, Nominal Mask Diameter, H rate, Row... Figure 9: General Linear Model 3: Cavity Diameter vs. Horizontal Undercut, Etch Time, Nominal Mask Diameter, Row, and HF Concentration... Figure : Residual Plots from General Linear Model 3... 6 Figure : Etch Rate (DERC) Original Analysis... 8 Figure : Individual Distribution Identification for Etch Rate (DERC)... 9 Figure 3: Etch Rate (DERC) Updated Analysis... 9 Figure : Etch Rate (DERC) Updated Analysis Continued... Figure : Cu Stress 9 (MPa) Original Analysis... Figure 6: Cu Stress 9 (MPa) Johnson Transformation... Figure : Cu Stress 9 (MPa) Johnson Transformation Continued... Figure 8: Cu Stress 9 (MPa) Updated Analysis... 3 Figure 9: Leak Rate (CH) Original Analysis... Figure : Leak Rate (CH) Box Cox Transformation... Figure : Leak Rate (CH) Updated Analysis... Figure : Average Sheet Rho (Ti) Original Analysis... 6 Figure 3: Average Sheet Rho (Ti) Johnson Transformation... 6 Figure : Average Sheet Rho (Ti) Updated Analysis... Figure : Etch Rate Coupon Johnson Transformation... Figure 6: Etch Rate Coupon Analysis... 8 Figure : Original Data with New Control Limits... 9 Figure 8: Original Data with New Control Limits... 9 Figure 9: Autocorrelation Plots... 3 Figure 3: Shear Strength... 33 Figure 3: A COIST Wafer... 3 Figure 3: Distribution of Shear Strength... 3 Figure 33: General Linear Model... 36 Figure 3: General Linear Model continued... 36 Figure 3: Tukey Method - Pairwise Comparisons among Levels of Location... 3 Figure 36: Tukey Method - Pairwise Comparisons among Levels of Wafer... 3 Figure 3: Wafer Map with Average Shear Strength... 38 Figure 38: ANOVA table... 39 Figure 39: Tukey Method - Pairwise Comparisons among Levels of Location... Figure : Tukey Method - Pairwise Comparisons among Levels of Wafer... Figure : Wafer Map with Average Shear Strength... Figure : Gage R & R Study using Xbar R method... 3 Figure 3: ANOVA Method Results Filmetrics F,. microns of SiO on Si... Figure : Tencor Step Height... Figure : Gage R & R Study -. um Feature... Figure 6: Gage R & R Study -. um Feature 8 Figure : Gage R & R Study. um Feature...

Figure 8: Gage R & R Study Angstroms of SiO on Si... Figure 9: Example slides of Minitab Presentation... Figure : Minitab s Assistant feature... Tables Table : Johnson Transformations... Table : Comparison of Original and New Control Limits... 3 Table 3: Classification of Measurement Tools (based on Part-to-Part variation)... 3 Table : Classification Comparison...

Introduction This Master s capstone was conducted in conjunction with Draper Laboratory, a non-profit research and development organization in Cambridge, Massachusetts. During a three month period, I worked for the Microfabrication Department at Draper, assisting with projects related to statistics and quality control. This report summarizes five major statistical studies that I completed. The first project dealt with an etch process - small holes, referred to as cavities, were etched into a glass wafer. Cavity diameters were recorded, as well as values for other parameters that were adjusted from wafer to wafer. For example, the length of time wafers spent in the etching bath varied. The goal was to see which variables potentially affected cavity diameter size, and to see if there were interesting relationships among variables. I fit general linear models to the data, with cavity diameter as the response and the other parameters as predictor variables. The final model, even though it included all significant variables, did not fit well and can be improved upon. In the second project, I examined process variables pertaining to a tool in the Microfabrication Laboratory. Control charts and capability ratios had been previously created for these variables. However, the charts and ratios used were incorrect because the assumption of normality was violated. I corrected the charts by first transforming the data to obtain normality, and then created new charts with the transformed data. I also found updated capability ratios and examined autocorrelation plots. The third project looked at the shear strength of tiny bumps on a wafer s surface. Engineers wanted to maximize shear strength, as well as see if strength was greatest at the center of a wafer. I fit general linear models for two sets of data, to see which wafers and locations produced the highest average shear strength. I also used Tukey s Method to examine pairwise comparisons. The fourth project was entirely devoted to Gage R & R studies. I reviewed data and results from Gage R & R studies completed in May of 3. I used Minitab to validate the results, as well as to analyze the data using a different method than previously used. I also conducted and analyzed new Gage R & R studies for three measuring tools in the Microfabrication Laboratory. Finally, I created PowerPoint presentations that explain how to use Minitab to perform data analysis. Specially, I picked statistical topics I thought were most useful and applicable in the department, and then showed how to execute them in Minitab. Other Draper employees can refer to the presentations in the future. The hope is that employees will utilize Minitab more often for analysis. Each chapter of this report focuses on one of the five projects. The chapters include a brief introduction with background information, goals of the statistical analysis, a description of the data, and statistical methods used to analyze the data. The chapters also include conclusions, comments, and recommendations. Minitab was the primary statistical software used to perform the data analysis.

Chapter : Etch Process Introduction The first analysis I completed dealt with an etch process. First, a mask is placed on top of a glass wafer. This mask, by the end of placement, has circular holes in it. The wafer (with the mask) is put inside a heated chemical bath of hydrofluoric acid. This is where etching took place. The etch process is believed to be isotropic, meaning material was removed in all directions at an equal rate. After a few hours, a hemispherical cavity forms in the glass at the sites where there was an opening in the mask. Figure displays the aforementioned process (the steps describing the placement of the mask are not included). Various parameters, such as etch time, were adjusted throughout the project. Engineers hoped to achieve nearly perfect holes of a specific size. Figure : Etch Process Goals The main goal for this study was to perform exploratory data analysis. I was asked to see what certain variables potentially affected cavity diameter size, and to see if there were interesting trends and correlations among variables.

Data/Variables There were 3 wafers included in the dataset. For most wafers, the same mask diameter size was used across the whole wafer. A few wafers had varying mask diameter sizes. The number of holes measured on each wafer varied. Response (Y) Cavity Diameter: size of hole, in µm Additional Variables/Possible Predictors (X, X, etc.) Wafer: identifier of the wafer Row, Column: Row and column of a cavity on the wafer Cavity Center X, Cavity Center Y, Mask Center X, Mask Center Y: X and Y coordinates, in µm Mask Diameter: Actual size of the mask hole, in µm Nominal Mask Diameter: The target mask hole diameter, in µm Mask Diameter Error: Difference between the actual and target mask hole diameter Etch Time: Total time wafer spent in the chemical bath, in hours Horizontal Undercut: Length of the overhang shown in step 9 of Figure, in µm H rate (µm/hr): Horizontal etch rate, equal to Horizontal Undercut / Etch Time HF concentration: Hydrofluoric acid % in the etching chemical bath Off Centering X, Off Centering Y: Difference between Cavity Center and Mask Center coordinates, in µm Off Centering D: Euclidean distance, in µm Off Centering Angle: Measured in degrees There were other variables in the data that I decided to exclude from my analysis, for various reasons. Some wafers contained information pertaining to cavity radius, cavity area, and cavity perimeter. These particular variables are directly related to cavity diameter, so, I could essentially use any of them as the response variable. However, cavity diameter was measured for every wafer, whereas the other variables were not. Therefore, it seemed reasonable to choose cavity diameter as the response. Judgment variables were also recorded- these kept track of defects at each cavity location. The location was deemed either GOOD ( n for some wafers) or BAD ( y ). As was the case with cavity radius, area, etc., not all judgment variables were listed for every wafer- in fact, very few wafers contained values for all judgment variables. In addition, I looked at a correlation matrix across all possible variables. I found the judgment variables to be significantly uncorrelated with any of the other variables, and thus, I decided to exclude them in the subsequent analysis. Statistical Analysis I first looked at the distribution of the response variable, cavity diameter. Many analyses assume (or require) that the response is normally distributed, so I wanted to check for that. The Anderson-Darling Normality Test, as well as Figures and 3 below indicate that cavity diameter is not normally distributed, and is skewed to the left.

Percent Figure : Distribution of Cavity Diameter Summary for Cavity Diameter A nderson-darling Normality Test A -Squared. P-V alue <. Mean 3. StDev 9. V ariance.6 Skew ness -.9 Kurtosis.8 N 8 9 3 Minimum 8. st Q uartile 3.9 Median 36.6 3rd Q uartile 386. Maximum 6. 9% C onfidence Interv al for Mean 96.9 3. 9% C onfidence Interv al for Median 3.8 3.3 9% Confidence Intervals 9% C onfidence Interv al for StDev.. Mean Median 3 3 3 36 38 Figure 3: Normal Quantile Plot of Cavity Diameter 99.99 99 9 Probability Plot of Cavity Diameter Normal - 9% CI Mean 3 StDev 9. N 8 AD.3 P-Value <. 8. 8 Cavity Diameter 6 8

To try and achieve normality, I proceeded to use a Box-Cox transformation, which raises the variable to a power. Using Minitab, I was able to find an optimal power. However, even after transformation, the response was still not normally distributed. Because of this, I had to be cautious with any interpretations I made from my analysis. The normality assumption was violated, so the validity of certain results are questionable. After looking at the response variable, I proceeded to examine the other variables as well. I looked at a correlation matrix and found strong evidence that cavity diameter was correlated with etch time and mask diameter. I also saw evidence of multicolinearity (correlation among predictor variables). In several cases, this multicolinearity made sense. For example, H-rate was correlated with Horizontal Undercut, but this is expected, since H-rate = Horizontal Undercut / Etch Time. However, I needed to take multicolinearity into consideration later on in my analysis, when I fit models to the data. Multicolinearity doesn t inhibit inference about the mean response, but it does result in large standard errors of the least squares estimates (model coefficients). It can also affect interpretations of the model coefficients. For a particular predictor variable, the model coefficient is the change in mean response when the predictor variable increases by unit, given the other variables remain fixed. If multicolinearity exists, the fixed assumption does not hold. After my initial first look, I decided to next try and fit a model to the data. If the model fit well, it could potentially be used as a reference in future work. One could attempt to hit a target cavity diameter, as long as the values chosen for the predictor variables were within the scope of the model (to avoid extrapolation). Fitting a model to the data would also reveal the significant predictors of cavity diameter, which may be useful information. I first attempted to fit a general linear model (specifically, a multiple linear regression model) to the data. where, is the response vector is a matrix containing observations for the p predictor variables is a vector of model coefficients is a vector of error (noise) terms As mentioned before, the first assumption (of normality) is violated, so some results from any general linear model (GLM) are questionable. For my first model, I chose to try cavity diameter versus all other variables, just to have a base. The results are shown in Figures and. Although the R-square value is equal to, other things indicate that the model is not a good fit. The residuals do not follow a normal distribution, and there are very large VIF values, which indicate multicolinearity. In addition, coefficients for two variables could not be estimated. Minitab has the option of performing a Box-Cox Transformation when fitting a model. As Figure shows, the optimal power is equal to (the transformed data is equal to itself). Finally, the lack of fit test indicates that the proposed model does not fit well.

Figure : General Linear Model : Cavity Diameter vs. all other variables * NOTE * Mask Diameter Error cannot be estimated and has been removed. * NOTE * Off-Centering X (μm) cannot be estimated and has been removed. Box-Cox transformation of the response with rounded lambda = The 9% CI for lambda is (.98,.) Regression Equation Cavity Diameter = -.86 +.39 Row -.8 Column -.38 Cavity Center X + 6. Cavity Center Y +.3698 Mask Center X - 6.3 Mask Center Y +.9688 Mask Diameter +.3 Etch Time +.38 Nominal Mask Dia. +.98 Horizontal Undercut (H) +.6 H rate (μm/hr) + 6.6 Off-Centering Y (μm) -.3893 Off-Centering D (μm) + 3.38e- Off-Centering Angle ( ) cases used, 36 cases contain missing values Coefficients Term Coef SE Coef T P VIF Constant -.86. -.899.369 Row.. 3...3E+ Column -.3.8 -.3..98E+ Cavity Center X -.3.3 -.3. 3.E+ Cavity Center Y 6. 6.66..9.83E+ Mask Center X.3.3.6.8 3.63E+ Mask Center Y -6.3 6.66 -..9 6.9839E+ Mask Diameter.96.9.396. 3.E+ Etch Time.. 9.. 6.6E+ Nominal Mask Dia...9 6.38. 3.99E+ Horizontal Undercut (H).99. 9.883. 3.89999E+ H rate (μm/hr).... 8.93E+ Off-Centering Y (μm) 6.6 6.66..9 6.368E+ Off-Centering D (μm) -.. -.98.3.6E+ Off-Centering Angle ( )...8.93.E+ Summary of Model S =.886933 R-Sq =.% R-Sq(adj) =.% PRESS = 63.6 R-Sq(pred) = 99.99% Analysis of Variance Source DF Seq SS Adj SS Adj MS F P Regression 996 996 8 83. Row 666 9 9. Column 36.3 Cavity Center X 8.688 Cavity Center Y 6.9 Mask Center X 3.86 Mask Center Y 8.9 Mask Diameter 393 838 838 8. Etch Time 93 6 6 83. Nominal Mask Dia. 689 3 3. Horizontal Undercut (H) 688 6 6 8. H rate (μm/hr) 8.3 Off-Centering Y (μm).9 Off-Centering D (μm).39 Off-Centering Angle ( ).936 Error 6 Lack-of-Fit 66 * * Pure Error Total 933

Frequency Deleted Residual Percent Deleted Residual Figure : Residual Plots from General Linear Model Residual Plots for Cavity Diameter Normal Probability Plot Versus Fits 99.99 99 9 - - -. - - - Deleted Residual Fitted Value 6 Histogram Versus Order 3 - - - -8 - - -9-6 -3 Deleted Residual 3 3 6 8 Observation Order 9 Next, I decided to try stepwise regression in Minitab, which utilizes forward and backward elimination simultaneously (for choosing the vital/significant predictor variables.) Variables were added to the model if its corresponding F-statistic was greater than or equal to, and removed if the F-statistic dropped below. F- statistics greater than or equal to means the p-value will be very significant (close to ). Figure 6 displays the results from the stepwise regression. After determining the significant variables in stepwise regression, I fit another general linear model. The results are shown in Figures and 8. Unfortunately, the model did not fit well. The residuals were still not normally distributed and VIF s were still very high. However, all of the predictor variables had significant p-values, as expected, and the R-square value was still high. The R-square value was almost equal to after adding two variables via stepwise regression. At the time of the analysis, I was still learning about the etch process. I didn t want to exclude a variable that may important or of interest to the engineers. This was why I kept all the significant variables in the second general linear model.

Frequency Deleted Residual Percent Deleted Residual Figure 6: Stepwise Regression with both Forward and Backward Elimination Stepwise Regression: Cavity Diameter versus ALL F-to-Enter: F-to-Remove: Response is Cavity Diameter on 6 predictors, with N = Step 3 6 Constant....3 -.9 -. Horizontal Undercut (H).3.98.999.9968.993.998 T-Value 6.9. 36. 36..8.8 P-Value...... Mask Diameter.9898.988.9.9336.98 T-Value 6.8 6.38. 8.3.9 P-Value..... Etch Time...6.39 T-Value 3.9 9.3 9.3 9.3 P-Value.... Nominal Mask Dia..398.96. T-Value.8.63 6.3 P-Value... H rate (μm/hr).. T-Value.68.69 P-Value.. Row. T-Value 3.6 P-Value. S 63.8.3.99.9.89.88 R-Sq. 99.99 99.99 99.99.. R-Sq(adj).3 99.99 99.99 99.99.. Figure : Residual Plots from General Linear Model Residual Plots for Cavity Diameter Normal Probability Plot Versus Fits 99.99 99 9 - - -. - - - Deleted Residual Fitted Value 6 Histogram Versus Order 3 - - - -8 - - -9-6 -3 Deleted Residual 3 3 6 8 Observation Order 9

Figure 8: General Linear Model : Cavity Diameter vs. Horizontal Undercut, Mask Diameter, Etch Time, Nominal Mask Diameter, H rate, Row General Regression Analysis: Cavity Diameter versus Horizontal Undercut, Mask Diameter, Etch Time, Nominal Mask Diameter, H rate, Row Box-Cox transformation of the response with rounded lambda = The 9% CI for lambda is (.9883,.83) Regression Equation Cavity Diameter = -.998 +.999 Horizontal Undercut (H) +.998 Mask Diameter +.39 Etch Time +.366 Nominal Mask Dia. +.8 H rate (μm/hr) +.636 Row 6 cases used, 3 cases contain missing values Coefficients Term Coef SE Coef T P VIF Constant -.999.838 -.3.96 Horizontal Undercut (H).999.3 3.86. 38.68 Mask Diameter.998.8.3. 86.6 Etch Time.39.3 9.6. 6.88 Nominal Mask Dia..3.89 6.9. 93. H rate (μm/hr).8.39.3. 8. Row.6.636 3.9..9 Summary of Model S =.8969 R-Sq =.% R-Sq(adj) =.% PRESS = 6.98 R-Sq(pred) = 99.99% Analysis of Variance Source DF Seq SS Adj SS Adj MS F P Regression 6 663 663 939 683. Horizontal Undercut (H) 9 8996. Mask Diameter 3 99 99 8. Etch Time 68 6 6 86. Nominal Mask Dia. 9 9 38. H rate (μm/hr). Row.8 Error 6 8 8 Lack-of-Fit 8 8 * * Pure Error 3 Total 66 3 Nominal mask diameter and mask diameter are two variables in GLM. They contain the same information, so I decided to remove one of them from the model (I kept nominal mask diameter). Also, horizontal undercut, etch time, and H rate were all in the model. Recall, H rate = Horizontal Undercut / Etch Time. I decided to also remove H-rate, to see if that alleviated the multicolinearity issues. At this time in my analysis, I was provided with information regarding HF concentration. It was believed that this would be an important variable, so I included it in the model. The results can be found in Figures 9 and. The results for GLM3 are much better than GLM and GLM. The predictor variables are significant, VIF values are all less than (indicates low multicolinearity) and the R-Squared value is close to as well. The residuals are slightly closer to being normally distributed, however they are still not (skewed to the left). However, the lack of fit test indicates the proposed model is not a good fit. I was told that the relationship between etch time and cavity diameter is not linear, so the results from the lack of fit test seem to make sense. Therefore, a different model is needed, such as a GLM with higher order terms, a generalized linear model (GLMM) with higher order terms, or a nonlinear model.

It s interesting to note that both HF concentration and etch time have negative coefficients and the corresponding confidence intervals, not shown, do not contain. This would mean that, if all other variables are held constant (which is possible), the average cavity diameter would decrease when either of these variables increases. This is not expected, at least to me it makes more sense that the etched holes would become larger the longer the wafer sat in the chemical bath or the stronger the concentration of HF. Since the lack of fit test indicates a linear model is not the correct choice, it s possible these coefficients are incorrect. Figure 9: General Linear Model 3: Cavity Diameter vs. Horizontal Undercut, Etch Time, Nominal Mask Diameter, Row, and HF Concentration General Regression Analysis: Cavity Diameter versus Horizontal Undercut, Etch Time, Nominal Mask Diameter, Row and HF concentration Box-Cox transformation of the response with rounded lambda =. The 9% CI for lambda is (.6,.) Regression Equation Cavity Diameter^. = -.6 +.8 Horizontal Undercut (H) -.669 Etch Time +.99 Row +.3 Nominal Mask Dia. -.939 HF Concentration (%) 83 cases used, 33 cases contain missing values Coefficients Term Coef SE Coef T P VIF Constant -.6 9.96 -.93. Horizontal Undercut (H).6. 89.3..93 Etch Time -.66.9 -...3 Row.9.99.83..93 Nominal Mask Dia...8 36.6.. HF Concentration (%) -.9.939-3.63..8 Summary of Model S = 8.6 R-Sq = 99.9% R-Sq(adj) = 99.9% PRESS =.3 R-Sq(pred) = 99.9% Analysis of Variance Source DF Seq SS Adj SS Adj MS F P Regression 8698 8698 3896 8. Horizontal Undercut (H) 89 68 68 6886. Etch Time 9 9.338 Row 89 3 3 3. Nominal Mask Dia. 96636 6368 6368 96. HF Concentration (%) 8 8 8 86. Error 9 9 6 Lack-of-Fit 3 3 38. Pure Error Total 8 8999

Frequency Residual Percent Residual Figure : Residual Plots from General Linear Model 3 Residual Plots for Cavity Diameter Normal Probability Plot Versus Fits 99.99 99 9 - -. - - Residual Fitted Value 8 3 Histogram Versus Order - - -3 - - Residual - - 3 Observation Order 6 8 9 Comments/Recommendations As shown previously, cavity diameter appears bi-modal. One possible reason for this is that engineers were targeting different size holes. The etching of the wafers was just the first step of a much more involved project, and certain size holes were needed. Another reason could be that not every hole on every wafer was measured. Select holes were measured, so a true random sample was not taken. I would suggest looking further into the bimodality of cavity diameter, and seeing if wafers can be grouped. Correlation within wafers was something I did not have a chance to look at, but is something that should be examined.

Chapter : Denton Introduction After a layer of dielectric material is placed on a wafer, the wafer then goes through a series of other production steps, which include receiving layers of metal and conductive material. In the Microfabrication Laboratory at Draper, the equipment responsible for these steps, such as receiving metal, is the Denton. Another employee at Draper had created control charts for variables pertaining to the Denton. For example, each time the Denton performed etching, the etch rate was recorded. Individual and moving range charts were created for the data. Capability ratios were calculated for the variables as well. The charts and ratios are updated as new data is collected, and are used to make sure the Denton processes are in control and meeting specifications. For several of the variables, the data was not normally distributed. This meant that the corresponding control charts and capability ratios were potentially incorrect. Shewhart s I-MR charts and certain capability ratios assume that the data follow a normal distribution. Goals I decided to correct the control charts and capability ratios for the cases where the data was not normally distributed. I used data transformations to obtain normality, and then found new charts and ratios. The updated information would hopefully provide better insight as to whether or not the process was in control and meeting specifications. By comparing the before and after charts, one could see if the results from prior to the transformation were incorrect as well. I was also prepared to propose and set up alternate charts to use if transformations did not work, or if I thought another chart was more appropriate. Certain charts, such as EWMA charts, are robust for non-normality, and are used to detect small shifts. However, for every variable that was non-normal, the transformation worked, and I obtained normality. The I-MR chart also was the most appropriate, because in each case, individual measurements were taken, as opposed to samples. Statistical Analysis Figure displays the original analysis for Etch Rate (DERC). There are two primary processes (known as recipes) for which Etch Rate is recorded. These recipes have code names they are Ajowan and Ginger. I utilized Minitab s Individual Distribution Identification tool in order to see if I could transform the data to obtain normality. This tool also tests whether the data follows one of the common, non-normal distributions as well. If the data followed one of the common non-normal distributions, Minitab would have allowed me to create control charts and perform capability analysis based on that distribution. Based on all the probability plots drawn (more than what is shown in Figure ), a Box-Cox transformation to obtain normality was the best choice. AD stands for Anderson Darling, which tests whether the sample data was drawn from a given probability distribution. One does not reject the null hypothesis, which states that the data follows the given distribution, when the p-value is greater than..

Box-Cox transformations attempt to obtain normality by raising all the data to a power. For example, one transformation may be: Here, is the original data point and is the transformed data. One requirement is that for all i. Figures 3 and display the new analysis for Etch Rate (DERC), using the transformed data, which was obtained by raising the data to the power 3. The process appears to be out of control. Observation goes outside the 3σ limits. Also, there appears to be an upward trend starting around observation. The Ginger recipe is within the specification limits, but is off center. Ajowan is not meeting specifications. The capability ratios from after are larger than before, but provide essentially the same interpretation. Figure : Etch Rate (DERC) Original Analysis Capability Histograms of Etch Rate (DERC) by Recipe Recipe = Ajowan LSL USL Within Cp.88 CPL -. CPU.9 Cpk -. 9 6 3. 3... 6.. 8. Within Overall Overall Pp.63 PPL -. PPU.368 Ppk -. Cpm * Recipe = Ginger Within Cp.86 CPL. CPU 3.9 Cpk....... 3.. LSL. 6.. 8. USL Within Overall Overall Pp.8 PPL.88 PPU.6 Ppk.88 Cpm * Original LSL = 6 USL = 88

Percent Percent Percent Percent Figure : Individual Distribution Identification for Etch Rate (DERC) 99 9 Normal - 9% C I Probability Plot for Etch Rate (DERC) Normal - 9% C I 99 9 Goodness of Fit Test Normal A D =.363 P-V alue <. Box-C ox Transformation A D =.33 P-V alue =.33 99 6 Etch Rate (DERC) Lognormal - 9% C I 8 3 Etch Rate (DERC) A fter Box-C ox transformation (lambda = 3) 3-Parameter Lognormal - 9% C I 99 Lognormal A D =. P-V alue <. 3-Parameter Lognormal A D =.333 P-V alue = * 9 9 Etch Rate (DERC) 3 3 3 36 Etch Rate (DERC) - T hreshold Figure 3: Etch Rate (DERC) Updated Analysis Capability Histograms of Etch Rate (DERC) ^ 3 by Recipe Within Cp. CPL -.6 CPU.9 Cpk -.6..... - Recipe = Ajowan LSL 3 6 8 USL Within Overall Overall Pp.88 PPL -.8 PPU 3. Ppk -.8 Cpm * Recipe = Ginger Within Cp.683 CPL.9 CPU.6 Cpk.9 6.. 3... - LSL 3 6 8 USL Within Overall Overall Pp.3 PPL.9 PPU 3.8 Ppk.9 Cpm *

Values Moving Range Individual Value Figure : Etch Rate (DERC) Updated Analysis Continued Process Capability Sixpack of Etch Rate (DERC) ^ 3 I Chart UCL=36633 _ X=98 Capability Histogram LSL USL Specifications LSL 66 USL 68 9 3 9 33 3 LCL=9 36 8 6 Moving Range Chart UCL=3 Normal Prob Plot A D:.33, P:.33 MR=686 9 3 9 33 3 LCL= 3 3 Last Observations 3 Observation 3 Within StDev 63 Cp. Cpk. PPM 339.69 Capability Plot Within O v erall Specs Overall StDev 69 Pp. Ppk. Cpm * PPM 3866.3 Updated LSL = 6 3 = 66 USL = 88 3 = 68 I used the same approach (Individual Distribution Identification tool) for the other variables that were not normally distributed as well. These variables are Cu Stress 9 (MPa), Leak Rate (CH), Average Sheet Rho (Ti), and Etch Rate Coupon. Etch Rate Coupon did not have an original analysis, but specification limits were provided to me, so I was able to complete my own analysis. The results are shown in Figures through 6. The Leak Rate (CH) data obtained normality through a Box Cox transformation. Cu Stress 9 (MPa), Average Sheet Rho (Ti), and Etch Rate Coupon, on the other hand, obtained normality via another type of transformation: Johnson Transformation. Johnson Transformations do not require that your original data be positive (Cheshire, ). These transformations try to obtain normality by selecting one of three types of functions. The function types are known as families and are shown in Table (Cheshire, ). Again, is the original data, is the transformed data, and are parameters. These transformations are monotone, which means that relationships among variables and data points are preserved.

Values Moving Range Individual Value Table : Johnson Transformations Johnson Family Transformation Function SB SL SU Figure : Cu Stress 9 (MPa) Original Analysis Process Capability Sixpack of Cu Stress 9 [MPa] 8 3 9 I Chart 3 3 3 9 UCL=9. _ X=93. LCL=-8.9 LSL Capability Histogram - - USL 6 8 Specifications LSL - USL 8 Moving Range Chart Normal Prob Plot A D: 3.88, P: <. 3 9 3 3 3 9 UCL=8. MR=6. LCL= 3 6 9 Last Observations Capability Plot - Observation 6 Within StDev 6.39 Cp. Cpk.9 PPM. Within Overall Specs Overall StDev 9. Pp. Ppk.8 Cpm * PPM.3 Original LSL = - USL = 8

Percent Percent P-Value for AD test Percent Percent Figure 6: Cu Stress 9 (MPa) Johnson Transformation 99 9 8 Probability Plot for Cu Stress 9 [MPa] 3-Parameter Loglogistic - 9% CI Normal - 9% CI 99 9 8 Goodness of Fit Test 3-Parameter Loglogistic A D =. P-V alue = * Johnson Transformation A D =. P-V alue =.9 Cu Stress 9 [MPa] - Threshold - Cu Stress 9 [MPa] After Johnson transformation Figure : Cu Stress 9 (MPa) Johnson Transformation Continued Johnson Transformation for Cu Stress 9 [MPa] Probability Plot for Original Data 99 9 N AD 3.88 P-Value <....33 Select a T ransformation.. Ref P 3 6 9....6.8 Z Value (P-Value =. means <=.).. Probability Plot for T ransformed Data 99 9 N AD. P-Value.9 P-V alue for Best Fit:.9 Z for Best F it:.33 Best Transformation Ty pe: SU Transformation function equals.68 +.9 * A sinh( ( X -. ) /.93 ) -

Values Moving Range Individual Value In the updated analysis for Cu Stress 9 (MPa) (Figure 8), observation 3 is outside the control limits. If assignable cause can be found for this point, it can be removed, and the limits can be recalculated. When this is done, the process appears to be in control. The process is meeting specification limits (besides observation 3), and based on Cpk, the process yield is about 99.3%. The Cp and Cpk ratios are much lower in the updated analysis compared with the original. Figure 8: Cu Stress 9 (MPa) Updated Analysis Process Capability Sixpack of Cu Stress 9 Johnson Transform I Chart UCL=3. _ X=. LSL Capability Histogram USL Specifications LSL -3.663 USL 3.9-3 9 3 3 3 9 LCL=-3. -3 - - 3 Moving Range Chart UCL=3.698 Normal Prob Plot A D:., P:.9 MR=.3 3 9 3 3 3 9 LCL= -. -. -.9 3 Last Observations Observation Within StDev.3 Cp.9 Cpk. PPM 38.3 Capability Plot Within Overall Specs Overall StDev.6 Pp.3 Ppk.96 Cpm * PPM 38.66 Updated LSL = -3.663 USL = 3.9

Percent Percent Percent Percent Values Moving Range Individual Value Figure 9: Leak Rate (CH) Original Analysis Process Capability Sixpack of Leak Rate (CH).... 6 6 I Chart 6 3 36 Moving Range Chart 6 UCL=.83E-8 _ X=.83E-9 LCL=-.33E-8 Capability Histogram LSL USL -.E-8 -.E-8....6.8. Normal Prob Plot A D: 3., P: <. Specifications LSL. USL... 6 6 6 3 36 6 UCL=.E-8 MR=.8E-9 LCL= -.E-8...... Last Observations 3 3 Observation Within StDev 6.98e-9 Cp.9 Cpk.3 PPM 38.69 Capability Plot Within O v erall Specs Overall StDev.39e-8 Pp. Ppk.3 Cpm * PPM 38. Original LSL = USL = e-8 Figure : Leak Rate (CH) Box Cox Transformation 99 9 Normal - 9% C I Probability Plot for Leak Rate (CH) Normal - 9% C I 99 9 Goodness of Fit Test Normal A D = 3. P-V alue <. Box-C ox Transformation A D =.9 P-V alue =.99 -.E-8 99.. Leak Rate (CH ) Lognormal - 9% C I. 3 Leak Rate (CH) A fter Box-C ox transformation (lambda = -.) 3-Parameter Lognormal - 9% C I 99 Lognormal A D =.3 P-V alue <. 3-Parameter Lognormal A D = 3.9 P-V alue = * 9 9... Leak Rate (CH ).....3 Leak Rate (CH) - T hreshold

Values Moving Range Individual Value The updated analysis for Leak Rate (CH) is shown in Figure. There are two out of control points (on both the individuals and moving range chart), and an upward trend starting at observation. The process appears to be out of control. The process is also not meeting specifications, indicated by the low Cp value. Figure : Leak Rate (CH) Updated Analysis Process Capability Sixpack of Leak Rate (CH) ^-. 3 I Chart UCL=388 _ X=3 LSL Capability Histogram USL Specifications LSL. USL. 6 6 6 3 36 6 LCL=8 8 6 3 Moving Range Chart UCL=8 Normal Prob Plot A D:.9, P:.99 MR=83 6 6 6 3 36 6 LCL= 3 3 3 Last Observations 3 Observation Within StDev 99 Cp. Cpk -. PPM 99993.6 Capability Plot Within O v erall Specs Overall StDev 69 Pp. Ppk -. Cpm * PPM 9868.38 Updated LSL = USL =.36

Percent Percent P-Value for AD test Values Moving Range Individual Value Figure : Average Sheet Rho (Ti) Original Analysis Process Capability Sixpack of Avg. SR (Ti) 8 9 8 I Chart 3 6 6 3 8 UCL=.88 _ X=8.68 LCL=.88 Capability Histogram LSL 6 8 USL Specifications LSL.6 USL. Moving Range Chart UCL=3.3 Normal Prob Plot A D: 3., P: <. 9 8 3 6 6 3 8 MR=. LCL=.... 8 6 6 Last Observations 8 Observation 8 Within StDev.933 Cp.6 Cpk. PPM 98. Capability Plot Within Overall Specs Overall StDev. Pp.86 Ppk.8 Cpm * PPM 996. Original LSL =.6 USL =. Figure 3: Average Sheet Rho (Ti) Johnson Transformation Johnson Transformation for Avg. SR (Ti) Probability Plot for Original Data 99.9 99 9 N 86 AD 3. P-Value <..6. Select a T ransformation.......3.....6.8 Z Value (P-Value =. means <=.).. Ref P Probability Plot for T ransformed Data 99.9 99 9 N 86 AD.66 P-Value.68 P-V alue for Best Fit:.689 Z for Best F it:. Best Transformation Ty pe: SU Transformation function equals -.638 +.866 * A sinh( ( X - 8.898 ) /.96 ). -

Percent Percent P-Value for AD test Values Moving Range Individual Value Figure shows the updated analysis for Average Sheet Rho (Ti). If assignable cause can be found for the one out of control point on the MR chart, then the process is in control. The process has a yield of about 96%. Figure : Average Sheet Rho (Ti) Updated Analysis Process Capability Sixpack of Avg. SR (Ti) Johnson Transform I Chart UCL=.688 _ X=.3 Capability Histogram LSL USL Specifications LSL -.868 USL.3-9 8 3 6 6 3 8 LCL=-.69 - - 3. Moving Range Chart UCL=3.6 Normal Prob Plot A D:.66, P:.68. MR=.998. 9 8 3 6 6 3 8 LCL= - Last Observations Capability Plot. -. -.9 6 Observation 8 8 Within StDev.886 Cp.8 Cpk. PPM 88.9 Within O v erall Specs Overall StDev.99 Pp. Ppk. Cpm * PPM 336. Updated LSL = -.868 USL =.3 Figure : Etch Rate Coupon Johnson Transformation Johnson Transformation for Etch Rate Coupon Probability Plot for Original Data 99.9 99 9 N 8 AD.388 P-Value <..8.6 Select a T ransformation.8. 6 8 9.....6.8 Z Value (P-Value =. means <=.).. Ref P Probability Plot for T ransformed Data 99.9 99 9 N 8 AD. P-Value.86 P-V alue for Best Fit:.869 Z for Best F it:.8 Best Transformation Ty pe: SU Transformation function equals -.368 +.9 * A sinh( ( X - 9.6 ) / 3.6 ). -

Values Moving Range Individual Value Figure 6 displays the analysis for Etch Rate Coupon. The process yield is about 99.3%. There are a couple of out of control points (one on the individuals chart, two on the MR chart). The variation between observations seems to decrease starting around observation 9 on the MR chart (many points below the centerline). It s possible this phenomenon could be explained by some assignable cause (such as replacement of new machine parts, etc.). If so, then the process would be in control from observation 9 onward. Figure 6: Etch Rate Coupon Analysis 3 Process Capability Sixpack of ER Coupon Johnson Transform I Chart UCL=.8 _ X=.8 LSL Capability Histogram USL Specifications LSL -3.3998 USL.38-3 LCL=-.8 9 33 9 6 3-3 - - Moving Range Chart UCL=3. Normal Prob Plot A D:., P:.86 MR=. 9 33 9 6 3 LCL= -.. -. 6 Last Observations 6 Observation 8 Within StDev.99 Cp. Cpk.89 PPM 38.3 Capability Plot Within Overall Specs Overall StDev. Pp.9 Ppk.8 Cpm * PPM 88.9 Comments Original: LSL = 6 USL = 96 Updated: LSL = -3.3998 USL =.38 One of my supervisors pointed out that in the original control chart for Leak Rate (CH), there are three points that are clearly out of control. In the updated charts, two points are still out of control, but a third point is within the control limits (it is on the edge of being out of control, however). This same supervisor said that the original chart is more useful for him, as it clearly displays the out of control points. He would be more likely to take action after seeing the original chart than after seeing my updated one. He said that in some cases, he is not concerned with false positives. I suggested that in the future, both charts (original and updated using transformed data) could be considered simultaneously. Although the original charts may give false positives, it s also possible that they give false negatives. Also, some employees may have difficulty interpreting the updated charts, so being able to look at the original chart may prove to be helpful. Another option is to take an updated control chart, and transform it back into original units. The individual data points will be exactly the same as the original, un-transformed data. The only difference is that there will be new control limits, which reflect the data s distribution. Figures and 6 show this for Etch Rate (DERC) and Leak Rate (CH). Table compares the original and new control limits.

Moving Range Individual Value Moving Range Individual Value Figure : Original Data with New Control Limits I-MR Chart of Etch Rate (DERC) - New Limits 8 6. _ X=6.9 9 3 Observation 9 33 3. 6 6.9 3 MR=.88 9 3 Observation 9 33 3 Figure 8: Original Data with New Control Limits I-MR Chart of Leak Rate (CH) - New Limits.E-.E-8.E-8.E-8.E+ 3.68E-8 _ X=.83E-9 8.86E- 6 6 6 Observation 3 36 6.E-.E-8.E-8.E-8.E+ 3.8E-8 MR=.8E-9 6 6 6 Observation 3 36 6

Autocorrelation Table : Comparison of Original and New Control Limits Etch Rate (Derc) Original Individual New Individual Original Moving Range New Moving Range LCL 3.9. UCL.86.. 6.9 Leak Rate (CH) Original Individual New Individual Original Moving Range New Moving Range LCL -.33E-8 8.86E- UCL.83E-8 3.68E-8.E-8 3.8E-8 Finally, I checked the data for autocorrelation. The results are shown in Figure 9. There appears to be significant correlation for observations two time units apart for Etch Rate (Derc) and for Average Sheet Rho (Ti). Also, there is significant correlation for observations two time units and nineteen units apart for Etch Rate Coupon. The significant estimates out of all 8 is not out of line with the expected number of significant sample autocorrelations when there is no process autocorrelation. Figure 9: Autocorrelation Plots a.) Etch Rate (DERC)..8.6 Autocorrelation Function for Etch Rate (DERC) (with % significance limits for the autocorrelations)... -. -. -.6 -.8 -. 3 Lag 6 8 9

Autocorrelation Autocorrelation b.) Cu Stress 9 (MPa)..8.6 Autocorrelation Function for Cu Stress 9 [MPa] (with % significance limits for the autocorrelations)... -. -. -.6 -.8 -. 3 6 8 Lag 9 3 c.) Average Sheet Rho (Ti)..8.6 Autocorrelation Function for Avg. SR (Ti) (with % significance limits for the autocorrelations)... -. -. -.6 -.8 -. 6 8 Lag 6 8

Autocorrelation Autocorrelation d.) Leak Rate (CH)..8.6 Autocorrelation Function for Leak Rate (CH) (with % significance limits for the autocorrelations)... -. -. -.6 -.8 -. 3 6 Lag 8 9 e.) Etch Rate Coupon..8.6 Autocorrelation Function for Etch Rate Coupon (with % significance limits for the autocorrelations)... -. -. -.6 -.8 -. 6 8 Lag 6 8

Chapter 3: COIST Wafers Part Introduction A collection of wafers, identified as COIST wafers, have tiny bumps on their surfaces. The shear strength, that is, the maximum shear stress which a material can withstand without rupture was tested for these bumps (Corrosionpedia, ). Figure 3: Shear Strength http://www.masterbond.com/sites/default/files/lpimages/physical-strength_properties-shear.png The COIST wafers had been produced using different methods/processes. This meant it was likely that the shear strength of the bumps varied across the set of wafers. It was also hypothesized that the shear strength varied within wafers, and that the bumps with the strongest shear strength were located at the center of the wafer. Goals The first goal of this analysis was to determine if shear strength varied across wafers. Ultimately, engineers were interested in finding out which specific process yielded the largest shear strengths. The second goal of the analysis was to see how much shear strength varied within a wafer. Data/Variables Wafers COIST 3 COIST 6 COIST COIST 8 COIST 9 COIST 6 COIST 6 The wafers were the same size (mm) and each wafer contained the same number of bumps. Bumps were organized in groups, and these groups were aligned in rows and columns. Figure 3 gives a rough picture of a COIST wafer. Eleven groups were selected by engineers (not random) and the total force at failure was measured for 8 randomly selected bumps in each group. Then, the shear strength was calculated. This was done for each of the seven wafers.

Figure 3: A COIST Wafer A Group of Bumps Row 3 Column (R3C) 8 bumps from a group tested Note: This is not an exact representation of a COIST wafer. There are more groups of bumps, as well as more bumps within a group than what is shown above. Response (Y) Shear Normalized: This is the shear strength I refer to throughout the rest of the analysis. Additional Variables/Possible Predictors (X, X, etc.) Location: Row # and column # of a group of bumps Shear Raw Data: Total Force at Failure Force Resisting Area: Size of the bump Statistical Analysis I first looked at the distribution of all wafers together variable (see Figure 3). It did not follow a normal distribution and was skewed to the left. It did not follow one of the common non-normal distributions either. I tried obtaining normality using a transformation, but neither a Box Cox transformation nor a Johnson Transformation were successful. This seems to makes sense: the distribution is possibly bi-modal, and monotone transformations (like Box Cox or Johnson) would not fix bi-modality. The reason for bimodality is unknown. I examined data for individual wafers, and almost all wafers contributed to the left tail of the distribution, as opposed to one or two wafers. I didn t see any common characteristics

between points within each peak. The shearing process requires an engineer to properly align the tool to each bump. It s possible the bimodality is a result of poor alignment. It s also possible multiple engineers sheared the same wafer. Figure 3: Distribution of Shear Strength Summary for Shear Normalized A nderson-darling Normality Test A -Squared 3. P-V alue <. Mean 8.669 StDev.386 V ariance.386 Skew ness -.988 Kurtosis -.66 N 6 6 8 9 Minimum.699 st Q uartile.96 Median 9. 3rd Q uartile 9.83 Maximum.6 9% C onfidence Interv al for Mean 8.33 8.999 9% C onfidence Interv al for Median 8.6 9. 9% Confidence Intervals 9% C onfidence Interv al for StDev.3.969 Mean Median 8. 8. 8.8 9. 9.6 I proceeded with my analysis, despite the lack of normality. I just needed to be careful with interpretations of some of the results. I decided to fit a general linear model to the data. I used wafer and location as categorical predictor variables and shear normalized (shear strength) as the response. The results are shown in Figures 33 and 3. Due to lack of time, interactions were not examined, but should be considered in future analyses. The constant coefficient is the overall shear strength for COIST 6 at location R9C (Row 9 Column ). The wafer and location coefficients are the change in average shear strength when referring to a particular wafer and location. For example, the average shear strength for any location on COIST 3 is.86 units larger than COIST 6 at location R9C. The average shear strength will increase by an additional. if referring to location RC6 on COIST 3. The model appears to be a relatively good fit. The residuals roughly follow a normal distribution, minus a few outliers. There doesn t seem to be any pattern in the Residual versus Fitted Value plot. All but one p-value are significant, indicating a difference in average shear strength between wafers and between locations. The R- square value is 8.6%, which is relatively good.

Frequency Deleted Residual Percent Deleted Residual Figure 33: General Linear Model General Regression Analysis: Shear Normalized versus Wafer, Location Coefficients Term Coef SE Coef T P 9% CI Constant 8.8.9.36. ( 8.698, 9.883) Wafer COIST - 3.86.96 3.88. (.36,.368) COIST - 6.9.99.36. (.39,.6863) COIST - -.6.38 -.63. (-.366,.33) COIST - 8.96.389.33. (.69683,.8) COIST - 9-3.6.636-9.88. (-3.633, -.989) COIST - 6.936.3 6.96. (.69,.6) Location RC -.398.696 -.996. (-.86, -.9369) RC6..3 3.9. (.6,.9) RC -.996.3-6.99. (-.983, -.693) RC -.688.39 -.. (-.8, -.93) RC -.88.638 -.9.6 (-.8, -.39) RC6.3.39.38. (.639,.8633) RC -.93.39 -.68.9 (-.39, -.98) RC6.68.638 3.68. (.338,.) R8C6.663.39.38. (.3,.383) R9C -..39-6.936. (-.69, -.3) Summary of Model S =.68 R-Sq = 8.6% R-Sq(adj) =.6% PRESS =.69 R-Sq(pred) = 6.6% Analysis of Variance Source DF Seq SS Adj SS Adj MS F P Regression 6 6.936 6.936.6.3 Wafer 6 3.996 3.38.386 83. Location 6.9 6.9 6.9 3.38 Error 33 8.39 8.39.6 Total 39 88. Figure 3: General Linear Model continued Residual Plots for Shear Normalized Normal Probability Plot Versus Fits 99.9 99 9 -. - - Deleted Residual - 6 8 Fitted Value Histogram Versus Order 3-3 - - Deleted Residual 3-8 6 - Observation Order 6 8 6 8 3 3

I next used Tukey s Method (for simultaneous confidence intervals) to group locations by average shear strength. Locations that do not share a letter are significantly different. The results are shown in Figure 3. I also grouped wafers using Tukey s Method as well (Figure 36). Figure 3: Tukey Method - Pairwise Comparisons among Levels of Location Location Mean Grouping RC6.9 A R8C6.3 A B RC6.93 A B C R9C 9.9 B C RC6 9. C RC 9.8 C D RC 8.9 C D RC 8.3 D RC 8.86 D R9C.6 D E RC 6.9 E Tukey 9% Simultaneous Confidence Intervals Individual confidence level = 99.86% Figure 36: Tukey Method - Pairwise Comparisons among Levels of Wafer Wafer Mean Grouping COIST - 8 9.89 A COIST - 6 9.3 A COIST - 3 9.88 A COIST - 6 9.8 A COIST - 6 9.6 A COIST - 8.6 B COIST - 9.3 C Tukey 9% Simultaneous Confidence Intervals Individual confidence level = 99.66% Conclusions Based on the above analysis, COIST 8 and COIST 6 had the greatest average shear strength compared to the other wafers in the analysis. However, only two other wafers are significantly different from COIST 8 and COIST 6. Location RC6 (Row Column 6) had the greatest shear strength out of all the locations. It seems, at first, that bumps at the center of the wafer have the strongest shear strength. Looking closer at the Tukey groupings and considering the rest the model, this isn t necessarily the case. Location RC6 is not significantly different from R8C6 and RC6. RC6 is located at the edge of the wafer, while the other two are near the center. Also, RC6 and R8C6 are right next to each other at the center of the wafer, but they are significantly different based on Tukey. See Figure 3. The general linear model suggests that COIST 6 at location RC6 will lead to the greatest shear strength.

Figure 3: Wafer Map with Average Shear Strength

Part Introduction In the first analysis, all wafers had the same number of bumps and the same locations were measured on each. I was also given data for wafers where this wasn t the case. Wafers had different numbers of bumps on them and a different sample size was taken for each wafer. The same locations weren t measured on every wafer. All wafers were the same size however, so, the ID for a location on one wafer corresponded to the same location on another wafer. Goals The goals of this analysis were to see which wafers and locations yielded the largest shear strength, and to examine how shear strength varied across a wafer. Statistical Analysis An ANOVA table (see Figure 38) shows that both wafer and location are significant. More variation occurs between wafers rather than within wafers. Only 6.3% of the variation is explained by those two factors, however. Figure 38: ANOVA table General Linear Model: Shear Normalized versus Wafer, Location Factor Type Levels Values Wafer fixed COIST - 9, COIST - 3, COIST - 3, COIST - 3, COIST - 3, COIST - 38, COIST - 39, COIST -, COIST -, COIST -, COIST - 3 Location fixed 9 A, G, A, B, D, G, C, D, D, E, 6C, 6E, 6F, F, G, H, 8A, 8F, 8G Analysis of Variance for Shear Normalized, using Adjusted SS for Tests Source DF Seq SS Adj SS Adj MS F P Wafer 33.63 33.3 3.333 8.. Location 8 39. 39. 3.36.. Error.83.83. Total 3 89.9 S =.886 R-Sq = 6.3% R-Sq(adj) =.% Figure 39 and Figure show the Tukey groups by location and by wafer. When referring to location, the rows are identified by a number (Row,, etc.) and the columns are identified by a letter (Column A, B, etc.). Location A is row column. Groups that do not share a letter are significantly different.

Figure 39: Tukey Method - Pairwise Comparisons among Levels of Location Location N Mean StDev Grouping (Tukey) H 8.866.8 A C 8.86.6 A F 8.68. A A 8..989 A B 6F 8.9. A B 8F 6.69.8 A B B 38.3.9 A B G 3.9.63 A B C E 6..63 A B C 8A..6 A B C G.9. A B C 6E 3.6. A B C 8G.8.83 A B C D.38.6 A B C A 6.699.6 B C D 3.6.3 B C 6C 6.8.9 B C D D 8 9.. C D G 8 8..839 D Tukey 9% Simultaneous Confidence Intervals Individual confidence level = 99.9% Figure : Tukey Method - Pairwise Comparisons among Levels of Wafer Wafer N Mean Grouping (Tukey) COIST - 3 8.8 A COIST 9 88.3 A B COIST 6.6 B C COIST 3 8.9 C D COIST 39 8.9 D E COIST 38 3. D E F COIST 8.68 E F COIST 3.6 E F COIST 3.6 F COIST 3 3.6 F COIST - 3 9.99 G Tukey 9% Simultaneous Confidence Intervals Individual confidence level = 99.8% Some locations only have 8 observations these observations were all taken from only one wafer. H: COIST C: COIST 9 F: COIST 3 A: COIST 6F: COIST 3 D: COIST G: COIST 9 COIST 9 had one of the largest overall average shear strengths. But, location G was measured only on this wafer, and it was the location with the lowest average shear strength. This location was not as low as the entire COIST 3 wafer however, which was consistently lower at each location compared with the other wafers. COIST 3 consistently had higher shear strengths at each location. Other wafers, such as COIST and COIST 3, had higher variation between locations. Locations at the center of the wafer do have large shear strengths, but there are also some equivalently high strengths near the edges (See Figure ).

Figure : Wafer Map with Average Shear Strength

Chapter : Gage R & R Studies Part : Review of Past Studies Introduction In May of 3, Gage R & R studies were conducted for three pieces of measuring equipment- Filmetrics F Filmetrics F Woollam Specifically, these tools measure the thickness of a thin film on a wafer. Past Procedure For each of the measurement tools above, three Gage R & R studies were conducted:. Three operators measured silicon wafers (parts) containing. microns of silicon dioxide. Each part was measured three times.. Three operators measured silicon wafers (parts) containing. microns of silicon dioxide. Each part was measured three times. 3. Three operators measured silicon wafers (parts) containing. microns of silicon dioxide. Each part was measured three times. Parts were measured in a random order, and not consecutively. In other words, an operator measured each part once, in a random order. The parts were measured a second time, again in random order. Finally, the parts, in a random order, were measured a third time. The Xbar R method was used to analyze the data collected. The Xbar R method is limited, however, because it does not include an operator part interaction term. In some cases, this interaction is significant, and should be taken into account. In addition, as shown in Figure, the Xbar R method can produce negative variance components. Other methods for analyzing Gage R & R studies avoid such issues. Restricted maximum likelihood, for example, does not produce negative components. Minitab has two options for analyzing a Gage R & R study: the ANOVA method and the Xbar R method. The ANOVA method fits a general linear model to calculate the variance components. It also includes the operator part interaction term in the variance components model, if it s deemed significant. If the interaction term is not significant, Minitab will fit a reduced model to calculate variance components. Occasionally, one or more of the estimated variance components could be less than zero when the operator part interaction term is insignificant (Minitab, ). If a variance component is negative, then Gage R&R will report it as zero (Minitab, ).

Figure : Gage R & R Study using Xbar R method MEMS Filmetrics F ---.µm SiO on Si Process: Thin Film Measurement Metrology Tool: MEMS Filmetrics Product Code Name: Gauge Study Box Parameter: SiO on Si Tool Make/Model: F Measurement Units: Å Specification +/-: Measurement Range: Gunnar Operator A Cheryl Operator B Jeff Operator C Part Average Sample Trial # Trial # Trial #3 Range Trial # Trial # Trial #3 Range Trial # Trial # Trial #3 Range..8. 9....8 3.. 9.. 8. 8. 6. 9. 3. 6. 8.. 3. 9. 3.8 96..9. 6. 3. 9. 9...3 83.. 3.. 83. 88.3 8.3... 96. 6.3 93.9 89. 3. 3. 9. 68.8 86.6.3 83.6. 3. 6.3 6.9 9.. 3. 6. 3. 99.6 6. 3.. Totals 6.6 6.8 68. 69. 63. 6. 63. 9.6 639. Part Average Range Average Range Op A 9.9 Average Range Op B 3. Average Range Op C 9.9 6. Average Test-Retest Range.6 Test for Control/Upper Range Limit.3 Repeatability (test-retest metric) 6. Reproducibility (Operator variance metric) -. 9. R&R (Repeatability and Reproducibility) Product variance metric 3.3 Total Variance 98. Classification Repeatability Proportion. Classification Class - captures >9% of product signal - Reproducibility Proportion. Use for Process Monitoring (i.e. SPC) or Production Class - captures -9% of product signal - Product Variation Proportion. Good for Production use, limited use for Process Monitoring Class 3 - captures -% of the product signal - Total Variance Proportions. Review tolerance and criticality for Production use Class - captures <% of the product signal - Probable error. Review tolerance and criticality for Production use This is the median error of the measurements, which limits the resolving ability of the measurement. Goals I decided to re-analyze the Gage R & R study data from May 3 using Minitab. I ran the data through both the Xbar R and ANOVA methods. First, I used the Xbar R method to see if I could duplicate the same results as the original. Then, I used the ANOVA method to see if the interaction term was significant, and if the results were much different than the Xbar R method. The original analysis classified each tool on a scale from to. This classification was based on the percentage of part-to-part variation (labeled as product variation percent in these studies). See Table 3. I also classified the tools based on the part-to-part variation from Minitab s output from the Xbar R and ANOVA methods. Table 3: Classification of Measurement Tools (based on Part-to-Part variation) Classification Class - captures >9% of product signal - Use for Process Monitoring (i.e. SPC) or Production Class - captures -9% of product signal - Good for Production use, limited use for Process Monitoring Class 3 - captures -% of the product signal - Review tolerance and criticality for Production use Class - captures <% of the product signal - Review tolerance and criticality for Production use

Statistical Analysis Minitab s Xbar R method produced similar results as the original method used. However, the results were not identical. Minitab s Xbar R method does not output negative variation - it sets negative variation equal to zero. This explains the slight differences in results. The method produced the same classification as the original method (see Table ). Table : Classification Comparison Tool Filmetrics F Filmetrics F Woolam SE Sample Type Original Product Variation % Original Xbar R Classification ANOVA Product Variation % ANOVA Classification Xbar R Product Variation % Minitab s Xbar R Classification.um SiO on Si % 6.6% 3.%.um SiO on Si % %.66%.um SiO on Si % % 6.%.um SiO on Si 93% 93.6% 9.8%.um SiO on Si 3% 8.6%.%.um SiO on Si 8%.9% 8.8%.um SiO on Si % 99.% 99.%.um SiO on Si 93% 8.99% 9.83%.um SiO on Si 8% 88.63% 88.9% In eight of the nine Gage R & R studies, the operator part interaction was not significant, according to the ANOVA method. All p-values were greater than.. Figure 3 shows the Minitab output for one of the studies. The only case where the interaction term was significant was for the Woollam, measuring. microns of SiO. The ANOVA method classified three of the studies differently than the original method. Those studies are highlighted in yellow in Table above. For these studies, the operator part interaction term was not significant, and thus not included in the variance components model. The differences in classification, therefore, might be due to the fact that the Xbar R method uses the range to estimate variation, whereas the ANOVA method uses sums of squares. In the case where the interaction term was significant (Woollam,. microns of SiO), the classification was the same across all three methods. The classifications for the Woollam studies were similar the tool is good to use for production use, as well as statistical process control. The Filmetrics F, on the other hand, is an inconsistent measuring tool. The classifications for the Filmetrics F studies were very different. The tool performed inconsistently for. microns of silicon dioxide, but was good for. microns and. microns.

Sample Mean Average Sample Range Percent Figure 3: ANOVA Method Results Filmetrics F,. microns of SiO on Si Gage R&R Study - ANOVA Method Gage name: Filmetrics F,. um SiO on Si Two-Way ANOVA Table With Interaction Source DF SS MS F P Wafer (Part).96... Operator.89..3.3 Wafer (Part) * Operator 8.89.36.366.69 Repeatability 3.8. Total.39 Alpha to remove interaction term =. Two-Way ANOVA Table Without Interaction Source DF SS MS F P Wafer (Part).96. 8.3. Operator.89..68.3 Repeatability 38.6.9 Total.39 Gage R&R %Contribution Source VarComp (of VarComp) Total Gage R&R.3 3.39 Repeatability.9 3.9 Reproducibility.. Operator.. Part-To-Part. 6.6 Total Variation.8. Process tolerance =. Study Var %Study Var %Tolerance Source StdDev (SD) (6 * SD) (%SV) (SV/Toler) Total Gage R&R.3.33 9.9. Repeatability.8.3 8.8. Reproducibility.39.98.96.96 Operator.39.98.96.96 Part-To-Part.38.68 8.38.3 Total Variation.9.3. 8. Number of Distinct Categories = Gage R&R (ANOVA) for. um Gage name: Date of study : F ilmetrics F,. um of SiO on Si Reported by : Tolerance: Misc: 8 Components of Variation % Contribution % Study Var.. um by Wafer (Part)..8 Gage R&R Repeat Reprod Part-to-Part R Chart by Operator Cheryl Gunnar Jeff UCL=... 3 Wafer (Part). um by Operator. _ R=.6.... 3 3 Wafer (Part) Xbar Chart by Operator Cheryl Gunnar Jeff 3 LCL= UCL=. _ X=.39 LCL=.86... Cheryl Gunnar Operator Wafer (Part) * Operator Interaction Jeff Operator Cheryl Gunnar Jeff. 3 3 Wafer (Part) 3. 3 Wafer (Part)

Comments/Suggestions I would suggest repeating the Filmetrics F Gage R & R studies specifically, the. microns of SiO. It would be interesting to see if the low classification (compared to the other two) was an anomaly. Minitab s ANOVA method can output a value known as %Tolerance. The %Tolerance compares the estimates of variation with the allowable spread of variation. The ANOVA method also outputs the range of values contained within 6 standard deviations (of the sample). One of my supervisors noted that these two pieces of information are more important to him than the actual Gage R & R variation percent (%Contribution of VarComp). The Gage R & R variation component is good when the goal is process improvement (Minitab, ). If the main interest is in how well the tool measures parts relative to specification limits, then %Tolerance is better to use (Minitab, ). Typically, in the Microfabrication Department at Draper, parts are allowed to vary as such: So, for a target of micron, the lower specification limit would be.8 microns, and the upper specification limit would be. microns. In all of the Gage R & R studies, the samples were well within the specification limits. Looking at Figure, the Gage R & R contribution to total variation is 3.39%. This value indicates the tool is unacceptable, with regards to process improvement. Greater than 3% is considered unacceptable, -3% is acceptable depending on the situation, and less than % is acceptable (Minitab, ). However, the %Tolerance is.%, which says the tool is acceptable with regards to the specifications. Part : New Studies Goals I was asked to conduct Gage R & R studies for other measurement tools in the Microfabrication Laboratory. The goal was to see if these other tools take consistent measurements, or if the measurement system variation is unacceptable. Process tolerance was considered when determining whether a tool was acceptable or unacceptable. In the case of these specific studies, if %Tolerance < 3%, the tool was classified as acceptable. Tencor The Tencor is a profilometer, a tool used to scan a wafer s surface. A stylus moves along the surface of the wafer. Any ups or downs the stylus makes (because of features on the wafer) are recorded. At the end of the scan, one can see the roughness of the surface, as well as the height and width of certain features. See Figure c. The black arrows show the movement of the stylus, and the red arrows indicate the step height. For the Tencor, two separate Gage R & R studies were conducted. The first study measured the height of a surface feature that was microns high. The second study measured the height of a micron feature. The studies were conducted as such: Three operators measured identical parts on a wafer. Each part was measured three times, not consecutively. Parts were measured once, in a random order, then a second time (random order) and a third time (random order).

Figure : Tencor Step Height a.) One Module on a Wafer b.) Top View (of area in the red box in part a.) c.) Side View Note: Not drawn to scale. The black arrows in c.) Side View show how the stylus moves along the surface. The red arrows indicate the step height. The ANOVA method was used to analyze the data, and the results of the two studies are shown below in Figures and 6. Figure : Gage R & R Study -. um Feature Gage R&R Study - ANOVA Method Gage R&R for Height Gage name: Tencor micron Feature Tolerance: +- % Misc: LSL =.6 USL =. Target = Two-Way ANOVA Table With Interaction Source DF SS MS F P Part 9.3.9 6.6. Operator.9.389.66. Part * Operator 8.638.369 3.3638. Repeatability 6.69.6 Total 89.898 Alpha to remove interaction term =. Gage R&R %Contribution Source VarComp (of VarComp) Total Gage R&R.389 33.86 Repeatability.6.9 Reproducibility.. Operator.66.8 Operator*Part.88 9.9 Part-To-Part.63 66. Total Variation.93. Process tolerance =.8 Study Var %Study Var %Tolerance Source StdDev (SD) (6 * SD) (%SV) (SV/Toler) Total Gage R&R.6.6 8.9 3.8 Repeatability.3.633 3.3.8 Reproducibility.9.8 6.98.6 Operator.98.68 3. 8. Operator*Part.968. 3.8 6.9 Part-To-Part.6.38 8.3 8. Total Variation.39.89..6 Number of Distinct Categories =

Sample Mean Average Sample Range Percent Gage R&R (ANOVA) for Height Gage name: Date of study : Tencor micron F eature Reported by : Noelle Tolerance: +- % Misc: LSL =.6 USL =. Target = Components of Variation Height by Part 8 % Contribution % Study Var % Tolerance..9.....9.9 Gage R&R Repeat RC3 RC RC3 RC R3C R3C RC3 R C R C R 6C3 R C3 RC RC3 RC R3C R3C RC3 Part Reprod R Chart by Operator Noelle Brad Dan Xbar Chart by Operator Noelle Brad Dan RC3 RC RC3 RC R3C R3C RC3 R C R C R6C3 RC3 RC RC3 Part Part-to-Part RC RC R6C3 RC3 RC RC3 RC R3 C R3 C R C3 R C RC R6C3 RC R3C R3C RC3 RC RC R6C3 RC3 RC RC3 RC R3 C R3 C R C3 RC RC R6C3 UCL=.368 _ R=.69 LCL= _ UCL=.99 X=.99 LCL=.99.9..9.9..9.9 RC3 RC RC3 RC R3C R3C RC3 Part Noelle RC3 RC RC3 RC Height by Operator R3C R3C Part Brad Operator Part * Operator Interaction RC3 RC RC R6C3 RC RC R6C3 Dan Operator Noelle Brad Dan Figure 6: Gage R & R Study -. um Feature Gage R&R Study - ANOVA Method Gage R&R for Height Gage name: Tencor micron Feature Tolerance: +- % Misc: LSL = USL = 6 Target = Two-Way ANOVA Table With Interaction Source DF SS MS F P Part 9.39.336.3. Operator..8 3.688. Part * Operator 8.3.9 8.8. Repeatability 6.998.333 Total 89.368693 Alpha to remove interaction term =. Gage R&R %Contribution Source VarComp (of VarComp) Total Gage R&R.68 3. Repeatability.333.36 Reproducibility.36. Operator.6.8 Operator*Part.8 9. Part-To-Part.3 6.3 Total Variation.3. Process tolerance = Study Var %Study Var %Tolerance Source StdDev (SD) (6 * SD) (%SV) (SV/Toler) Total Gage R&R.8.9 6.98 3.6 Repeatability.3.36.3.3 Reproducibility.66.6393. 3. Operator.39.383.6. Operator*Part.933.6 3.89.8 Part-To-Part.66.86 8.8. Total Variation.669.6. 6.38 Number of Distinct Categories =

Sample Mean Average Sample Range Percent Gage R&R (ANOVA) for Height Gage name: Date of study : Tencor micron F eature Reported by : Noelle Tolerance: +- % Misc: LSL = USL = 6 Target = 8 Components of Variation % Contribution % Study Var. Height by Part % Tolerance.3.. Gage R&R Repeat Reprod R Chart by Operator Brad Dan Noelle Part-to-Part UCL=.3 _ R=.89.3..3 3 6 Part Height by Operator 8 9. LCL=..3 3 6 8 9 3 6 Part Brad Dan Noelle 8 9 Xbar Chart by Operator 3.3 3 6 8 9 3 6 Part 8 9 3 6 6 8 8 9 _ UCL=.396 X=.3 LCL=.33.3..3 Brad Dan Operator Part * Operator Interaction Noelle Operator Brad Dan Noelle 9.3 3 6 Part 8 9 In both cases, the operator part interaction was significant and thus included in the model. For the first study (. micron feature), the %Tolerance was 3.8%. Based on this, the measurement tool is acceptable. For the second study ( microns) the %Tolerance was 3.6%, indicating that the tool is acceptable. The Repeatability measure is larger in the first study than in the second. At the time, two of the operators were newer users of the Tencor. It s possible that some of the variation is due to unfamiliarity with the tool. The Reproducibility measure is larger in the second study than in the first. The particular parts measured in the second study were difficult to locate on the wafer. It s possible that the operators accidentally measured different parts from one another. If the study was aimed toward process improvement, the Gage R & R variance component would be considered unacceptable (>3%) in both cases. Considering both studies, I believe the Tencor takes acceptable measurements with respect to the specification limits.

Tencor P 6 The Tencor P 6 is another profilometer used to scan a wafer s surface. For this tool, only one Gage R & R study was performed. The height of a surface feature that was microns high was measured. The study was conducted as such: Three operators measured identical parts on a wafer. Each part was measured three times, not consecutively. Parts were measured once, in a random order, then a second time (random order) and a third time (random order). The ANOVA method was used to analyze the data and the results of the study is shown below in Figure. The tool is acceptable with respect to the specification limits, since the %Tolerance is less than 3%. The tool is also acceptable with respect to process improvement. Figure : Gage R & R Study. um Feature Gage R&R Study - ANOVA Method Gage R&R for Height Gage name: Tencor P-6 Misc: LSL =.6 USL =. Target = Two-Way ANOVA Table With Interaction Source DF SS MS F P Part 9.6.69 3.8. Operator.98.399.. Part * Operator 8.393.9.96. Repeatability 6.. Total 89.86 Alpha to remove interaction term =. Two-Way ANOVA Table Without Interaction Source DF SS MS F P Part 9.6.69 3.86. Operator.98.399 6.98. Repeatability 8.6. Total 89.86 Gage R&R %Contribution Source VarComp (of VarComp) Total Gage R&R. 3.36 Repeatability..8 Reproducibility..6 Operator..6 Part-To-Part.69 96.6 Total Variation.. Process tolerance =.8 Study Var %Study Var %Tolerance Source StdDev (SD) (6 * SD) (%SV) (SV/Toler) Total Gage R&R.9.9 8.3 3.68 Repeatability.9.68 6.3 3.36 Reproducibility.999.99.. Operator.999.99.. Part-To-Part.699.96 98.3 9. Total Variation.6.63..6 Number of Distinct Categories =

Sample Mean Average Sample Range Percent Gage R&R (ANOVA) for Height Gage name: Tencor P-6 Date of study : Reported by : Tolerance: Misc: LSL =.6 USL =. Target = Components of Variation % Contribution % Study Var. Height by Part % Tolerance.9.....9 Gage R&R 3 6 8 9 Repeat 3 6 Part Reprod R Chart by Operator Brad Dan Noelle 8 9 Xbar Chart by Operator Brad Dan Noelle 3 Part-to-Part 6 8 9 UCL=. _ R=.6 LCL= _ X=.9 UCL=.99 LCL=.9.9..9.9..9 3 6 Part Height by Operator Brad Dan Operator Part * Operator Interaction 8 9 Noelle Operator Brad Dan Noelle.9 3 6 8 9 3 6 Part 8 9 3 6 8 9.9 3 6 Part 8 9 Filmetrics F Filmetrics F measures the thickness of a thin film on a wafer. For this study, operators measured the thickness of a layer of silicon dioxide on a silicon wafer. More specifically, Three operators measured silicon wafers (parts) containing Angstroms (. micron) of silicon dioxide. Each part was measured three times, not consecutively. Parts were measured once, in a random order, then a second time (random order) and a third time (random order). The results of the study are shown in Figure 8. Specification limits were not provided for this study. I used Target ± %, in my analysis. Based on those limits, the tool is acceptable. If the tolerance was lowered to Angstroms (LSL = 99, USL = ), the tool would be unacceptable, and would have a %Tolerance of 3..

Figure 8: Gage R & R Study Angstroms of SiO on Si Gage R&R Study - ANOVA Method Gage R&R for Thickness Gage name: Filmetrics F Misc: LSL = 8 Target = USL = Two-Way ANOVA Table With Interaction Source DF SS MS F P Part 6. 6.6 9.8. Operator.9.96 8.93. Part * Operator 8.6.6.3.99 Repeatability 3 6.66.6 Total. Alpha to remove interaction term =. Two-Way ANOVA Table Without Interaction Source DF SS MS F P Part 6. 6.6.86. Operator.9.96.68. Repeatability 38 8.3.3 Total. Gage R&R %Contribution Source VarComp (of VarComp) Total Gage R&R.8363 8. Repeatability.3 39. Reproducibility.38 9.6 Operator.38 9.6 Part-To-Part.669. Total Variation 3.8. Process tolerance = Study Var %Study Var %Tolerance Source StdDev (SD) (6 * SD) (%SV) (SV/Toler) Total Gage R&R.8. 69.8.89 Repeatability. 6.6 6.6.69 Reproducibility.88 3.39 3..8 Operator.88 3.39 3..8 Part-To-Part.93.9.9.9 Total Variation.83.8.. Number of Distinct Categories =

Sample Mean Average Sample Range Percent Gage R&R (ANOVA) for Thickness Gage name: Date of study : Filmetrics F Reported by : Tolerance: Misc: LSL = 8 Target = USL = 8 Components of Variation % Contribution Thickness by Part % Study Var % Tolerance 6 8 6 Gage R&R Repeat 8 6 6 8 6 Part Reprod R Chart by Operator Brad Dan Noelle Xbar Chart by Operator Brad Dan Noelle 8 6 6 8 6 Part Part-to-Part 6 8 6 6 6 8 6 6 UCL=.63 _ R=.8 LCL= UCL=8.33 _ X=6.89 LCL=.6 8 6 Part Thickness by Operator 6 Brad Dan Operator Part * Operator Interaction 8 6 8 6 6 Part 6 Noelle Operator Brad Dan Noelle Summary Tool Tencor Tencor P 6 Filmetrics F Status Acceptable Acceptable Acceptable (depending on tolerance)

Chapter : Minitab PowerPoint Presentations Introduction The Microfabrication Department at Draper (Cambridge location) primarily uses Excel and MATLAB to perform its data analysis. Draper has a license for Minitab, although many people are unfamiliar with the program. One of my supervisors asked me to learn Minitab and then create PowerPoint presentations about it. Specifically, he wanted me to pick statistical methods I thought would be most useful and applicable in the department, and then show how to execute them in Minitab. Other employees could later refer to the presentations to find out which specific statistical test to choose and how to carry it out. The goal is that employees will utilize Minitab more often for data analysis. Data/Variables I used a variety of data for these presentations. I used Draper s data (some collected by myself, some from others), data I found online, and data from my previous courses as WPI. Presentations I created four presentations about Minitab:. Introduction/Basics a. Minitab Environment b. Uploading Data to Minitab/Saving Projects c. Graphical Summary d. Normality Test e. Descriptive Statistics f. Confidence Intervals and Plots g. ANOVA tables h. Hypothesis Tests i. One-Way ANOVA j. Tukey Comparisons k. Standardizing Data. Quality Control a. Control Charts i. X Bar and R chart ii. X Bar and S chart iii. Individuals and MR chart iv. EWMA chart v. CUSUM chart vi. P, C, and U chart vii. T chart b. Capability Analysis i. C P, C PL, C PU, C PK, C PM 3. Gage R & R Studies a. ANOVA vs. Xbar R methods b. Nested vs. Crossed

. Design of Experiments a. Terminology b. Factorial Designs i. Full Factorial Design ii. General Full Factorial Design iii. Fractional Factorial Design iv. Screening Experiments v. Replication vi. Blocking vii. Interaction Plots c. Split Plot Designs d. Response Surface Designs i. Central Composite Design ii. Box Behnken Design iii. Contour/Surface Plots iv. Optimization Figure 9 displays example pages of the presentation. Figure 9: Example slides of Minitab Presentation a.) Table for Determining What Control Chart to Use

b.) Parts of a Control Chart c.) Example of How to Set Up an Xbar R Chart

Conclusions/Critique I found Minitab to be a very easy and straightforward program to learn. For those unfamiliar with statistics, Minitab s Assistant feature will be very helpful. Assistant has the user follow a flow chart of questions to help them figure out what particular statistical test to choose. Assistant will also output interpretations and warnings for the user. Figure displays Assistants flow chart for control charts. However, take note that Assistant s flow charts often do not include all available options. Referring back to Figure, notice that EWMA, CUSUM, and C-charts are not listed in the flow chart, but they are available in the drop-down menu. In addition, while creating these presentations and using Minitab on a regular basis, I noticed some limitations with the program. There are many features that Minitab lacks, which other programs, such as SAS, do have. For example, SAS has the ability to fit generalized linear models, which are used for non-normal data. Minitab, to my knowledge, does not have this feature. Nevertheless, for the type of data that the Microfabrication Department deals with every day, I believe Minitab is a good program to use, and has almost all of the necessary features. Figure : Minitab s Assistant feature